Wire up semaphores on command buffer execution.
Change-Id: I90fe0eb53cba85d75efbcee9ef3fc5ebd6d8e173
diff --git a/src/intel/vulkan/anv_batch_chain.c b/src/intel/vulkan/anv_batch_chain.c
index 8ebe964..9279917 100644
--- a/src/intel/vulkan/anv_batch_chain.c
+++ b/src/intel/vulkan/anv_batch_chain.c
@@ -1148,9 +1148,9 @@
return true;
}
-VkResult
-anv_cmd_buffer_execbuf(struct anv_device *device,
- struct anv_cmd_buffer *cmd_buffer)
+VkResult anv_cmd_buffer_execbuf(struct anv_device* device, struct anv_cmd_buffer* cmd_buffer,
+ uint32_t wait_semaphore_count, anv_semaphore_t* wait_semaphores,
+ uint32_t signal_semaphore_count, anv_semaphore_t* signal_semaphores)
{
struct anv_batch *batch = &cmd_buffer->batch;
struct anv_block_pool *ss_pool =
@@ -1278,8 +1278,8 @@
cmd_buffer->surface_relocs.relocs[i].presumed_offset = -1;
}
- VkResult result = anv_device_execbuf(device, &execbuf.execbuf, execbuf.bos);
-
+ VkResult result = anv_device_execbuf(device, &execbuf.execbuf, execbuf.bos, wait_semaphore_count,
+ wait_semaphores, signal_semaphore_count, signal_semaphores);
anv_execbuf_finish(&execbuf, &cmd_buffer->pool->alloc);
return result;
diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index 2d9d1b5..af78d65 100644
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -830,7 +830,7 @@
execbuf.rsvd1 = device->context_id;
execbuf.rsvd2 = 0;
- result = anv_device_execbuf(device, &execbuf, exec_bos);
+ result = anv_device_execbuf(device, &execbuf, exec_bos, 0, NULL, 0, NULL);
if (result != VK_SUCCESS)
goto fail;
@@ -1107,12 +1107,13 @@
*pQueue = anv_queue_to_handle(&device->queue);
}
-VkResult
-anv_device_execbuf(struct anv_device *device,
- struct drm_i915_gem_execbuffer2 *execbuf,
- struct anv_bo **execbuf_bos)
+VkResult anv_device_execbuf(struct anv_device* device, struct drm_i915_gem_execbuffer2* execbuf,
+ struct anv_bo** execbuf_bos, uint32_t wait_semaphore_count,
+ anv_semaphore_t* wait_semaphores, uint32_t signal_semaphore_count,
+ anv_semaphore_t* signal_semaphores)
{
- int ret = anv_gem_execbuffer(device, execbuf);
+ int ret = anv_gem_execbuffer(device, execbuf, wait_semaphore_count, wait_semaphores,
+ signal_semaphore_count, signal_semaphores);
if (ret != 0) {
/* We don't know the real error. */
return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
@@ -1169,7 +1170,10 @@
pSubmits[i].pCommandBuffers[j]);
assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
- result = anv_cmd_buffer_execbuf(device, cmd_buffer);
+ result = anv_cmd_buffer_execbuf(device, cmd_buffer, pSubmits[i].waitSemaphoreCount,
+ (anv_semaphore_t*)pSubmits[i].pWaitSemaphores,
+ pSubmits[i].signalSemaphoreCount,
+ (anv_semaphore_t*)pSubmits[i].pSignalSemaphores);
if (result != VK_SUCCESS)
goto out;
}
@@ -1177,7 +1181,7 @@
if (fence) {
struct anv_bo *fence_bo = &fence->bo;
- result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
+ result = anv_device_execbuf(device, &fence->execbuf, &fence_bo, 0, NULL, 0, NULL);
if (result != VK_SUCCESS)
goto out;
diff --git a/src/intel/vulkan/anv_magma.cc b/src/intel/vulkan/anv_magma.cc
index d3819cb..e69a90b 100644
--- a/src/intel/vulkan/anv_magma.cc
+++ b/src/intel/vulkan/anv_magma.cc
@@ -109,14 +109,17 @@
return 0;
}
-int anv_gem_execbuffer(anv_device* device, drm_i915_gem_execbuffer2* execbuf)
+int anv_gem_execbuffer(anv_device* device, drm_i915_gem_execbuffer2* execbuf,
+ uint32_t wait_semaphore_count, anv_semaphore_t* wait_semaphores,
+ uint32_t signal_semaphore_count, anv_semaphore_t* signal_semaphores)
{
DLOG("anv_gem_execbuffer");
if (execbuf->buffer_count == 0)
return 0;
- uint64_t required_size = DrmCommandBuffer::RequiredSize(execbuf);
+ uint64_t required_size =
+ DrmCommandBuffer::RequiredSize(execbuf, wait_semaphore_count, signal_semaphore_count);
uint64_t allocated_size;
uint64_t cmd_buf_id;
@@ -135,7 +138,18 @@
return DRET_MSG(error, "magma_system_map failed");
}
- if (!DrmCommandBuffer::Translate(execbuf, cmd_buf_data)) {
+ std::vector<uint64_t> wait_semaphore_ids(wait_semaphore_count);
+ for (uint32_t i = 0; i < wait_semaphore_count; i++) {
+ wait_semaphore_ids[i] = magma_system_get_semaphore_id(wait_semaphores[i]);
+ }
+
+ std::vector<uint64_t> signal_semaphore_ids(signal_semaphore_count);
+ for (uint32_t i = 0; i < signal_semaphore_count; i++) {
+ signal_semaphore_ids[i] = magma_system_get_semaphore_id(signal_semaphores[i]);
+ }
+
+ if (!DrmCommandBuffer::Translate(execbuf, std::move(wait_semaphore_ids),
+ std::move(signal_semaphore_ids), cmd_buf_data)) {
error = magma_system_unmap(magma_connection(device), cmd_buf_id);
DASSERT(!error);
magma_system_free(magma_connection(device), cmd_buf_id);
diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index 89f4fa9..9d2fdf5 100644
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -603,9 +603,10 @@
void anv_device_init_blorp(struct anv_device *device);
void anv_device_finish_blorp(struct anv_device *device);
-VkResult anv_device_execbuf(struct anv_device *device,
- struct drm_i915_gem_execbuffer2 *execbuf,
- struct anv_bo **execbuf_bos);
+VkResult anv_device_execbuf(struct anv_device* device, struct drm_i915_gem_execbuffer2* execbuf,
+ struct anv_bo** execbuf_bos, uint32_t wait_semaphore_count,
+ anv_semaphore_t* wait_semaphores, uint32_t signal_semaphore_count,
+ anv_semaphore_t* signal_semaphores);
int anv_gem_connect(struct anv_device* device);
void anv_gem_disconnect(struct anv_device* device);
@@ -617,8 +618,9 @@
void anv_gem_close(struct anv_device* device, anv_buffer_handle_t gem_handle);
uint32_t anv_gem_userptr(struct anv_device *device, void *mem, size_t size);
int anv_gem_wait(struct anv_device* device, anv_buffer_handle_t gem_handle, int64_t* timeout_ns);
-int anv_gem_execbuffer(struct anv_device *device,
- struct drm_i915_gem_execbuffer2 *execbuf);
+int anv_gem_execbuffer(struct anv_device* device, struct drm_i915_gem_execbuffer2* execbuf,
+ uint32_t wait_semaphore_count, anv_semaphore_t* wait_semaphores,
+ uint32_t signal_semaphore_count, anv_semaphore_t* signal_semaphores);
int anv_gem_set_tiling(struct anv_device* device, anv_buffer_handle_t gem_handle, uint32_t stride,
uint32_t tiling);
int anv_gem_create_context(struct anv_device *device);
@@ -1225,8 +1227,10 @@
void anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary,
struct anv_cmd_buffer *secondary);
void anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer);
-VkResult anv_cmd_buffer_execbuf(struct anv_device *device,
- struct anv_cmd_buffer *cmd_buffer);
+VkResult anv_cmd_buffer_execbuf(struct anv_device* device, struct anv_cmd_buffer* cmd_buffer,
+ uint32_t wait_semaphore_count, anv_semaphore_t* wait_semaphores,
+ uint32_t signal_semaphore_count,
+ anv_semaphore_t* signal_semaphores);
VkResult anv_cmd_buffer_reset(struct anv_cmd_buffer *cmd_buffer);
diff --git a/src/magma/drm_command_buffer.cc b/src/magma/drm_command_buffer.cc
index d264412..97da6e3 100644
--- a/src/magma/drm_command_buffer.cc
+++ b/src/magma/drm_command_buffer.cc
@@ -6,9 +6,10 @@
#include "magma_system.h"
#include "magma_util/dlog.h"
#include "magma_util/macros.h"
-#include <inttypes.h>
-uint64_t DrmCommandBuffer::RequiredSize(drm_i915_gem_execbuffer2* execbuf)
+uint64_t DrmCommandBuffer::RequiredSize(drm_i915_gem_execbuffer2* execbuf,
+ uint32_t wait_semaphore_count,
+ uint32_t signal_semaphore_count)
{
auto execobjects = reinterpret_cast<drm_i915_gem_exec_object2*>(execbuf->buffers_ptr);
@@ -19,11 +20,16 @@
num_relocations += execobjects[res_index].relocation_count;
}
- return sizeof(magma_system_command_buffer) + sizeof(magma_system_exec_resource) * num_resources +
+ return sizeof(magma_system_command_buffer) +
+ (wait_semaphore_count + signal_semaphore_count) * sizeof(uint64_t) +
+ sizeof(magma_system_exec_resource) * num_resources +
sizeof(magma_system_relocation_entry) * num_relocations;
}
-bool DrmCommandBuffer::Translate(drm_i915_gem_execbuffer2* execbuf, void* command_buffer_out)
+bool DrmCommandBuffer::Translate(drm_i915_gem_execbuffer2* execbuf,
+ std::vector<uint64_t> wait_semaphore_ids,
+ std::vector<uint64_t> signal_semaphore_ids,
+ void* command_buffer_out)
{
DASSERT((execbuf->flags & I915_EXEC_HANDLE_LUT) != 0);
@@ -32,11 +38,21 @@
magma_system_command_buffer* command_buffer =
reinterpret_cast<magma_system_command_buffer*>(command_buffer_out);
- magma_system_exec_resource* exec_resources =
- reinterpret_cast<magma_system_exec_resource*>(command_buffer + 1);
+ uint64_t* dst_wait_semaphore_ids = reinterpret_cast<uint64_t*>(command_buffer + 1);
+ uint64_t* dst_signal_semaphore_ids =
+ reinterpret_cast<uint64_t*>(dst_wait_semaphore_ids + wait_semaphore_ids.size());
+ magma_system_exec_resource* exec_resources = reinterpret_cast<magma_system_exec_resource*>(
+ dst_signal_semaphore_ids + signal_semaphore_ids.size());
magma_system_relocation_entry* relocation_entries =
reinterpret_cast<magma_system_relocation_entry*>(exec_resources + num_resources);
+ for (uint32_t i = 0; i < wait_semaphore_ids.size(); i++) {
+ dst_wait_semaphore_ids[i] = wait_semaphore_ids[i];
+ }
+ for (uint32_t i = 0; i < signal_semaphore_ids.size(); i++) {
+ dst_signal_semaphore_ids[i] = signal_semaphore_ids[i];
+ }
+
uint32_t res_reloc_base = 0;
for (uint32_t res_index = 0; res_index < num_resources; res_index++) {
@@ -44,7 +60,8 @@
auto src_res = &execobjects[res_index];
auto src_res_relocs = reinterpret_cast<drm_i915_gem_relocation_entry*>(src_res->relocs_ptr);
- DLOG("translating res_index %u handle 0x%" PRIx64 " start_offset 0x%lx length 0x%lx",
+ DLOG("translating res_index %u handle 0x%" PRIx64 " start_offset 0x%" PRIx64
+ " length 0x%" PRIx64,
res_index, src_res->handle, src_res->rsvd1, src_res->rsvd2);
uint32_t num_relocations = dst_res->num_relocations = src_res->relocation_count;
@@ -80,6 +97,8 @@
command_buffer->num_resources = num_resources;
command_buffer->batch_buffer_resource_index = num_resources - 1; // by drm convention
command_buffer->batch_start_offset = execbuf->batch_start_offset;
+ command_buffer->wait_semaphore_count = wait_semaphore_ids.size();
+ command_buffer->signal_semaphore_count = signal_semaphore_ids.size();
return true;
}
diff --git a/src/magma/drm_command_buffer.h b/src/magma/drm_command_buffer.h
index dd2198b..e0e1823 100644
--- a/src/magma/drm_command_buffer.h
+++ b/src/magma/drm_command_buffer.h
@@ -6,17 +6,22 @@
#define DRM_COMMAND_BUFFER_H
#include "i915_drm.h"
+#include <vector>
class DrmCommandBuffer {
public:
// Returns the number of bytes needed for the magma_system_command_buffer
- // and the associated data structures for |execbuf|
- static uint64_t RequiredSize(drm_i915_gem_execbuffer2* execbuf);
+ // and the associated data structures for |execbuf| with |wait_semaphore_count| +
+ // |signal_semaphore_count| semaphores.
+ static uint64_t RequiredSize(drm_i915_gem_execbuffer2* execbuf, uint32_t wait_semaphore_count,
+ uint32_t signal_semaphore_count);
// Writes the magma_system_command_buffer and associated data structures
// into |command_buffer_out|. |command_buffer_out| must point to a buffer
- // that is at least RequiredSize(|exec_buf|) bytes
- static bool Translate(drm_i915_gem_execbuffer2* execbuf, void* command_buffer_out);
+ // that is sufficiently large, see RequiredSize.
+ static bool Translate(drm_i915_gem_execbuffer2* execbuf,
+ std::vector<uint64_t> wait_semaphore_ids,
+ std::vector<uint64_t> signal_semaphore_ids, void* command_buffer_out);
private:
DrmCommandBuffer() {}
diff --git a/tests/unit_tests/test_drm_command_buffer.cc b/tests/unit_tests/test_drm_command_buffer.cc
index 3b22712..4207351 100644
--- a/tests/unit_tests/test_drm_command_buffer.cc
+++ b/tests/unit_tests/test_drm_command_buffer.cc
@@ -38,16 +38,23 @@
.flags = I915_EXEC_HANDLE_LUT,
};
- uint64_t size = DrmCommandBuffer::RequiredSize(&execbuffer2);
- ASSERT_EQ(sizeof(magma_system_command_buffer), size);
+ std::vector<uint64_t> wait_semaphores;
+ std::vector<uint64_t> signal_semaphores;
+
+ uint64_t size = DrmCommandBuffer::RequiredSize(&execbuffer2, 0, 0);
+ EXPECT_EQ(sizeof(magma_system_command_buffer), size);
std::vector<uint8_t> buffer(size);
- EXPECT_TRUE(DrmCommandBuffer::Translate(&execbuffer2, buffer.data()));
+
+ EXPECT_TRUE(DrmCommandBuffer::Translate(&execbuffer2, wait_semaphores, signal_semaphores,
+ buffer.data()));
auto command_buffer = reinterpret_cast<magma_system_command_buffer*>(buffer.data());
EXPECT_EQ(-1, (int)command_buffer->batch_buffer_resource_index);
EXPECT_EQ(0u, command_buffer->batch_start_offset);
EXPECT_EQ(0u, command_buffer->num_resources);
+ EXPECT_EQ(0u, command_buffer->wait_semaphore_count);
+ EXPECT_EQ(0u, command_buffer->signal_semaphore_count);
}
std::unique_ptr<Buffer> CreateBuffer(uint64_t size)
@@ -58,13 +65,22 @@
return std::make_unique<Buffer>(connection_, handle, size);
}
- void WithBuffers(bool add_relocs)
+ void WithBuffers(bool add_relocs, uint32_t wait_semaphore_count, uint32_t signal_semaphore_count)
{
std::vector<std::unique_ptr<Buffer>> buffers;
buffers.push_back(CreateBuffer(PAGE_SIZE));
buffers.push_back(CreateBuffer(PAGE_SIZE));
+ std::vector<uint64_t> wait_semaphore_ids;
+ for (uint32_t i = 0; i < wait_semaphore_count; i++) {
+ wait_semaphore_ids.push_back(10 + i);
+ }
+ std::vector<uint64_t> signal_semaphore_ids;
+ for (uint32_t i = 0; i < signal_semaphore_count; i++) {
+ signal_semaphore_ids.push_back(100 + i);
+ }
+
std::vector<drm_i915_gem_relocation_entry> exec_relocs_0;
std::vector<drm_i915_gem_relocation_entry> exec_relocs_1;
std::vector<drm_i915_gem_exec_object2> exec_res;
@@ -120,23 +136,37 @@
.flags = I915_EXEC_HANDLE_LUT,
};
- uint64_t size = DrmCommandBuffer::RequiredSize(&exec_buffer);
- uint64_t expected_size = sizeof(magma_system_command_buffer) +
- sizeof(magma_system_exec_resource) * exec_res.size();
- if (add_relocs)
- expected_size +=
- sizeof(magma_system_relocation_entry) * (exec_relocs_0.size() + exec_relocs_1.size());
+ uint64_t size = DrmCommandBuffer::RequiredSize(&exec_buffer, wait_semaphore_ids.size(),
+ signal_semaphore_ids.size());
+ uint64_t expected_size =
+ sizeof(magma_system_command_buffer) +
+ (wait_semaphore_ids.size() + signal_semaphore_ids.size()) * sizeof(uint64_t) +
+ sizeof(magma_system_exec_resource) * exec_res.size() +
+ sizeof(magma_system_relocation_entry) * (exec_relocs_0.size() + exec_relocs_1.size());
EXPECT_EQ(expected_size, size);
std::vector<uint8_t> buffer(size);
- EXPECT_TRUE(DrmCommandBuffer::Translate(&exec_buffer, buffer.data()));
+ EXPECT_TRUE(DrmCommandBuffer::Translate(&exec_buffer, wait_semaphore_ids,
+ signal_semaphore_ids, buffer.data()));
auto command_buffer = reinterpret_cast<magma_system_command_buffer*>(buffer.data());
EXPECT_EQ(exec_buffer.buffer_count - 1, command_buffer->batch_buffer_resource_index);
EXPECT_EQ(exec_buffer.batch_start_offset, command_buffer->batch_start_offset);
EXPECT_EQ(exec_buffer.buffer_count, command_buffer->num_resources);
+ EXPECT_EQ(wait_semaphore_ids.size(), command_buffer->wait_semaphore_count);
+ EXPECT_EQ(signal_semaphore_ids.size(), command_buffer->signal_semaphore_count);
- auto exec_resource = reinterpret_cast<magma_system_exec_resource*>(command_buffer + 1);
+ auto semaphores = reinterpret_cast<uint64_t*>(command_buffer + 1);
+ for (uint32_t i = 0; i < wait_semaphore_count; i++) {
+ EXPECT_EQ(wait_semaphore_ids[i], semaphores[i]);
+ }
+ semaphores += wait_semaphore_count;
+ for (uint32_t i = 0; i < signal_semaphore_count; i++) {
+ EXPECT_EQ(signal_semaphore_ids[i], semaphores[i]);
+ }
+
+ auto exec_resource =
+ reinterpret_cast<magma_system_exec_resource*>(semaphores + signal_semaphore_count);
for (uint32_t i = 0; i < exec_res.size(); i++) {
EXPECT_EQ(exec_resource->buffer_id, buffers[i]->id());
EXPECT_EQ(exec_resource->offset, exec_res[i].rsvd1);
@@ -179,11 +209,11 @@
TEST(DrmCommandBuffer, SomeBuffers)
{
TestDrmCommandBuffer test;
- test.WithBuffers(false);
+ test.WithBuffers(false, 1, 2);
}
TEST(DrmCommandBuffer, BuffersWithRelocs)
{
TestDrmCommandBuffer test;
- test.WithBuffers(true);
+ test.WithBuffers(true, 3, 2);
}
